A long distance arc detection method

Abstract

The application discloses a long-distance electric arc detection method, and relates to the technical field of power system fault detection, and comprises the following steps: S100, long-distance electric arc detection model construction; S200, signal acquisition and preprocessing; S300, double-mode image generation; S400, long-distance electric arc detection model training; and S500, long-distance electric arc detection. The application solves the problem of electric arc detection under long-distance signal attenuation and provides a more accurate electric arc detection method.

Description

Technical Field

[0001] This invention relates to the field of power system fault detection technology, and in particular to a long-distance arc detection method. Background Technology

[0002] During the long-term operation of power systems, especially in long-distance power transmission scenarios, arc faults are more likely to occur due to long signal transmission distances and numerous environmental interference factors. They are also highly concealed and difficult to detect.

[0003] Traditional detection methods suffer from signal attenuation and noise interference, resulting in low accuracy in arc identification in complex industrial environments. Long-distance transmission lines are subject to complex electromagnetic noise interference. Because the noise interference signal has non-stationary and non-Gaussian characteristics and overlaps with the characteristic frequency band of the arc, it increases the difficulty of fault detection.

[0004] CN 120408456 A discloses an "AC arc detection method and related device based on current and harmonic fluctuations". It collects current signals at different sampling frequencies using two data acquisition devices. One high-frequency signal is used to extract harmonic features, and the other low-frequency signal is used to extract current waveform features. The two signals are input in parallel to a binary classification model and a fluctuation model. The arc detection result is determined by comparing the detection results of the two models. This method reduces the false alarm rate of arc by comparing the model results. However, it relies on two independent acquisition systems, which increases hardware cost and system complexity.

[0005] CN 112444707 A discloses an "arc detection based on frequency analysis". It measures the amplitude of the current signal in several preset, non-overlapping frequency ranges through multiple bandpass filters, and determines whether an arc has occurred based on these amplitude ranges or mathematical relationships. This method starts from the frequency domain and the circuit implementation is relatively simple, but it relies on a predefined and fixed arc model and discards the dynamic information in the time domain, resulting in insufficient utilization of feature information.

[0006] CN 120372212 A discloses a photovoltaic arc detection method and system based on neural networks. The method uses a hybrid architecture combining CNN and Transformer to detect arcs. CNN is used to extract local features from voltage time-series data, while Transformer captures long-term dependencies between sequences. This method improves the ability to capture time-series features through the CNN-Transformer hybrid model. However, it only uses voltage time-domain waveforms as input, ignoring the rich harmonics and high-frequency transient features exhibited by the arc in the frequency domain, and does not consider the signal attenuation problem during long-distance transmission.

[0007] Therefore, those skilled in the art are dedicated to developing a long-distance arc detection method. Summary of the Invention

[0008] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is that the prior art has low accuracy in detecting long-distance electric arcs.

[0009] This invention is based on dual-modal images. It deeply fuses high-frequency enhanced spectrograms and time-domain waveforms at the data level and uses them as a unified input for a single deep learning model. It does not rely on a preset or fixed arc data model. The deep learning model adaptively learns the time-frequency characteristics of the arc from the fused dual-modal images. It introduces frequency-domain modes and enhances the arc characteristics through adaptive attenuation compensation and high-frequency enhancement specificity.

[0010] In one embodiment of the present invention, a long-distance arc detection method is provided, comprising the following steps:

[0011] S100, Long-distance arc detection model construction: Based on deep learning mechanism, a long-distance arc detection model is constructed, and the state of the long-distance arc detection model is set to the training state.

[0012] S200, Signal Acquisition and Preprocessing: Raw current signals are acquired at key nodes of the power system. When the long-distance arc detection model is in the training state, the raw current signals are labeled with normal samples or arc samples. When the long-distance arc detection model is in the inference state, the raw current signals are not labeled. The raw current signals are standardized to obtain standardized current signals. The standardized current signals are then preprocessed to obtain preprocessed current signals.

[0013] S300, Dual-modal image generation: Draw a high-frequency enhanced spectrum diagram of the preprocessed current signal, extract key time-domain features, draw a time-domain waveform diagram marked with key time-domain features, and fuse the high-frequency enhanced spectrum diagram and the time-domain waveform diagram into a dual-modal image. When the long-distance arc detection model is in the training state, execute step S400; otherwise, execute step S500.

[0014] S400. Training the long-distance arc detection model: Receive dual-modal images as a dataset, dynamically adjust the learning rate using a cosine annealing strategy to complete the training of the long-distance arc detection model, set the state of the long-distance arc detection model to the inference state, and return to step S200.

[0015] S500: Long-distance arc detection, receiving dual-modal images, issuing an alarm when an arc fault is detected; when normal operation is detected, returning to step S200 to continue detection.

[0016] Optionally, in the long-distance arc detection method in the above embodiments, the long-distance arc detection model includes an initial convolutional layer, a backbone feature layer, an attention enhancement layer, a long-distance enhancement layer, and a decision classification layer. The initial convolutional layer extracts basic features, the backbone feature layer extracts high-level visual features of the basic features, the attention enhancement layer integrates an attention mechanism to adaptively weight the high-level visual features to obtain attention-enhanced features, the long-distance enhancement layer further enhances the long-distance signal transmission-related features through global pooling and fully connected layers to obtain long-distance feature-enhanced features, and the decision classification layer adopts a fully connected layer structure to transform the long-distance feature-enhanced features into decision classification to obtain normal / arc binary classification probabilities.

[0017] Furthermore, in the long-distance arc detection method of the above embodiments, step S100 includes:

[0018] S110. Construct the initial convolutional layer and extract basic features using a 3×3 convolutional kernel, as shown in the following formula:

[0019] ,

[0020] in, This represents the basic features output by the initial convolutional layer. This represents the model input, i.e., the bimodal image. Indicates the convolution kernel weights; Indicates the bias term;

[0021] S120. Construct the backbone feature layer using a lightweight convolutional neural network based on the MobileNet architecture. The basic features are processed through cascaded bottleneck blocks to obtain high-level visual features, as shown in the following formula:

[0022] ,

[0023] in, Indicates high-level visual features; Indicates the number of cascaded bottleneck blocks; Indicates cascading bottleneck block operations;

[0024] S130. Construct an attention enhancement layer to enhance high-level visual features and important features related to arc detection through a channel attention mechanism, i.e., attention-enhanced features, as shown in the following formula:

[0025] ,

[0026] in, Indicating attention enhancement features, This represents a multilayer perceptron. This represents the Sigmoid activation function. This indicates channel-by-channel multiplication. Indicates global average pooling. Indicates global max pooling;

[0027] S140. Construct a long-range enhancement layer. Capture the overall statistical information of the attention-enhanced features through global average pooling. Perform a non-linear transformation through two fully connected layers to extract the long-range feature enhancement features. The formula is as follows:

[0028] ,

[0029] in, Features that represent long-range feature enhancement; and These represent the weight matrices of the first transform fully connected layer and the second transform fully connected layer, respectively. and These represent the bias vectors of the first transform fully connected layer and the second transform fully connected layer, respectively. Represents the ReLU activation function; This represents the SoftMax activation function;

[0030] S150. Construct a decision classification layer, including a first classification layer and a second classification layer, to transform the long-distance feature enhancement features into decision classifications, namely the confidence scores for normal state or electric arc, as shown in the following formula:

[0031] ,

[0032] in, Indicates the confidence level of a normal state or an electric arc. Indicates the confidence level of the normal state. Indicates the confidence level of the electric arc; and These represent the weight matrices of the first and second classification layers, respectively. and These represent the bias vectors of the first and second classification layers, respectively.

[0033] S160. Update the model state and put the long-distance arc detection model into training state.

[0034] Furthermore, in the long-distance arc detection method in the above embodiments, the cascaded bottleneck block includes extended convolution, depthwise separable convolution, compression excitation module and projection layer. Extended convolution increases the number of channels in the input data to a higher dimension, depthwise separable convolution performs feature extraction, compression excitation module adaptively calibrates the channel feature response, and projection layer reduces the number of channels to the required output channels.

[0035] Furthermore, in the long-distance arc detection method in the above embodiments, the number of cascaded bottleneck blocks ranges from 12 to 20. For resource-constrained embedded devices and scenarios with high real-time requirements, fewer bottleneck blocks can be used, but this will reduce accuracy. For tasks with sufficient computing resources and complexity, the number of bottleneck blocks can be appropriately increased, but this will increase the detection time.

[0036] Preferably, in the long-distance arc detection method of the above embodiments, the number of cascaded bottleneck blocks is 16, which can fully extract features and balance accuracy and efficiency.

[0037] Optionally, in the long-distance arc detection method in any of the above embodiments, the key nodes of the power system include combiner boxes, inverters, line interfaces, important load access points, historically fault-prone sections, and the coordination boundaries of system protection devices.

[0038] Optionally, in the long-distance arc detection method of any of the above embodiments, step S200 includes:

[0039] S210. Acquisition of raw current signal: Acquire the raw current signal at a sampling frequency of [frequency missing]. The signal length is For exceeding The signal is truncated, and the insufficient signal is cut off. The signal is zero-filled to ensure uniform input signal length; when the long-distance arc detection model is in the training state, the original current signal carries the label of normal sample or arc sample, and when the long-distance arc detection model is in the inference state, the original current signal does not carry the label.

[0040] S220. Standardization processing: The zero-mean unit variance method is used to eliminate the effects of baseline drift and amplitude variation, resulting in a standardized current signal, as shown in the following formula:

[0041] ,

[0042] in, This represents the value of the standardized current signal. Indicates the first The value of the original current signal at each sampling point n Values ​​range from 0 to N -1 is a positive integer; The mean of the original current signal is represented by the following calculation method: ; The standard deviation of the original current signal is represented by the following method: ;

[0043] S230, Adaptive gain compensation: Based on signal strength, adaptive gain compensation is performed on the standardized current signal to compensate for signal attenuation. The formula is as follows:

[0044] ,

[0045] in, Indicates the first The current signal value after adaptive attenuation compensation at each standardized current signal sampling point; This indicates the compensation amplification factor, used to enhance weak arc signals; This represents the weak signal strength threshold, used to determine whether the original signal is too weak, and thus whether attenuation compensation needs to be triggered.

[0046] S240. Signal Enhancement: Gaussian white noise is added to the current signal after adaptive attenuation compensation. When the long-distance arc detection model is in the training state, the intensity of the Gaussian white noise is greater than 0; when the long-distance arc detection model is in the inference state, the intensity of the Gaussian white noise is 0. The preprocessed current signal is obtained, as shown in the following formula:

[0047] ,

[0048] in, The value of the current signal that represents noise enhancement, i.e., the pre-processed current signal; Let the noise sequence have a mean of 0 and a variance of . The normal distribution This represents the noise intensity coefficient, which achieves an acceptable balance between improving generalization ability and maintaining feature integrity.

[0049] Furthermore, in the long-distance arc detection method of any of the above embodiments, the normal sample is the current signal collected when the power system is operating normally, and the arc sample is the current signal collected when the power system generates an arc.

[0050] Optionally, in the long-distance arc detection method of any of the above embodiments, the ratio of normal samples to arc samples is 1:1.

[0051] Optionally, in the long-distance arc detection method of any of the above embodiments, the arc occurs in the form of an arc caused by insulation damage, loose connection, etc. during actual operation, or an arc generator triggering an arc in an experimental scenario.

[0052] Optionally, in the long-distance arc detection method of any of the above embodiments, the sampling frequency... The range is 100-500kHz.

[0053] Preferably, in the long-distance arc detection method of the above embodiments, the sampling frequency is... It is 250kHz.

[0054] Optionally, in the long-distance arc detection method of any of the above embodiments, the signal length The range is 512-2048 sampling points.

[0055] Preferably, in the long-distance arc detection method of the above embodiments, the signal length There are 1024 sampling points.

[0056] Optionally, in the long-distance arc detection method of any of the above embodiments, The value ranges from 1.5 to 3.5, with 1.5 for scenarios with slight signal attenuation and 3.5 for scenarios with severe signal attenuation.

[0057] Preferably, in the long-distance arc detection method of the above embodiments, .

[0058] Optionally, in the long-distance arc detection method of any of the above embodiments, a weak signal strength threshold is represented. The range is 0.02-0.1. When the noise level at the site is low and the requirement for early warning of weak electric arcs is high, the range should be biased towards the lower limit of 0.02. When the interference at the site is strong, the range should be biased towards the upper limit of 0.1.

[0059] Preferably, in the long-distance arc detection method of the above embodiments, .

[0060] Optionally, in the long-distance arc detection method of any of the above embodiments, the noise intensity coefficient The range is 0.02-0.10. If you want the model to focus more on learning the features of clear signals, you should appropriately bias towards the lower limit of the range (0.02); if you are dealing with a known high-noise environment, you should appropriately bias towards the upper limit of the range (0.10).

[0061] Preferably, in the long-distance arc detection method of the above embodiments, .

[0062] Optionally, in the long-distance arc detection method of any of the above embodiments, step S300 includes:

[0063] S310. Calculate the frequency domain distribution. Apply a short-time Fourier transform to the preprocessed current signal to calculate the spectrum values. The formula is as follows:

[0064] ,

[0065] in, Indicates time frame and frequency point Spectral value at; Indicates the length of the window; Indicates the time frame index. This represents the total number of time frames, calculated as follows: , This indicates frame shift, preferably 32 sampling points; Indicates frequency index, ,in, ; This represents the Hanning window function, centered at... ;

[0066] S320. Calculate the spectral energy. Square the spectral values ​​to calculate the spectral energy, and then convert the linear-scale spectral energy to the logarithmic scale using a logarithmic transformation, as shown in the following formula:

[0067] ,

[0068] in, This represents the spectral energy after the logarithm of the transformation, in decibels (dB). This represents a small constant added to avoid taking the logarithm of zero;

[0069] S330, Spectral Energy Enhancement: Enhances spectral energy in the frequency band where arc characteristics are prominent, highlighting the high-frequency features of the arc. The formula is as follows:

[0070] ,

[0071] ,

[0072] in, Indicates at frequency point The magnitude of the increase at that location; and These represent the lower and upper limits of the high-frequency enhancement range, respectively. Indicates the maximum gain; Indicates the center enhancement frequency; Indicates the enhancement parameter; It represents exponentiation with base e;

[0073] S340. Draw a high-frequency enhanced spectrum diagram. The high-frequency enhanced spectrum diagram is drawn in the form of a heatmap. The time axis and frequency axis are constructed as the horizontal and vertical axes, respectively, with units of milliseconds (ms) and kilohertz (kHz). The formula is as follows:

[0074] ,

[0075] ,

[0076] in, Indicates the first The actual time corresponding to each time frame; Frequency index The corresponding frequency, spectral energy After normalization, energy-color mapping is performed to the spectrogram using the Viridis color scale; key temporal features are extracted, and temporal waveforms marked with key temporal features are plotted.

[0077] S350. Extract key time-domain features: Extract key time-domain features from the preprocessed current signal, including peak position, signal RMS value, and waveform factor, as shown in the following formula:

[0078] ,

[0079] ,

[0080] ,

[0081] in, This indicates the peak position, which is the sampling position corresponding to the point where the absolute value of the signal amplitude is the largest; It represents the effective value of the signal, reflecting the average energy; This represents the waveform factor, which characterizes the peak characteristics of the signal;

[0082] S360. Draw a time-domain waveform. Based on key time-domain features, draw a time-domain waveform marked with key time-domain features, constructing a time axis as the horizontal axis, with the unit being milliseconds (ms). The formula is as follows:

[0083] ,

[0084] in Indicates the first The time value corresponding to each sampling point;

[0085] The vertical axis represents the noise-enhanced current signal value. Mark key feature parameters;

[0086] S370. Generate a dual-modal image by fusing the high-frequency enhanced spectrogram with the time-domain waveform marked with key time-domain features in a 1×2 sub-graph format to generate a dual-modal image, mining the time-frequency joint features of the signal, and improving the feature expression capability. When the long-distance arc detection model is in the training state, execute step S400; otherwise, execute step S500.

[0087] Furthermore, in the long-distance arc detection method of any of the above embodiments, the window length... The range is 64-256 sampling points. The short window range is 64-96 sampling points, which is suitable for rapidly changing transient arc detection. The medium window range is 96-160 sampling points, which is suitable for most arc detection application scenarios. The long window range is 160-256 sampling points, which is suitable for steady-state arc detection with relatively low time accuracy.

[0088] Preferably, in the long-distance arc detection method of the above embodiments, the window length... There are 128 sampling points.

[0089] Furthermore, in the long-distance arc detection method of the above embodiments, a small constant is added. Scope arrive ;

[0090] Preferably, in the long-distance arc detection method of the above embodiments, .

[0091] Furthermore, in the long-distance arc detection method of the above embodiments, the lower limit of the high-frequency enhancement range... The range is 15 kHz to 30 kHz, which determines the starting point for capturing low-frequency harmonic components in the arc. Below 15 kHz, there are too many power frequency harmonics and background noise, which dilute the arc characteristic components. Above 30 kHz, the low-frequency characteristics in the initial stage of the arc are filtered out too early, resulting in information loss.

[0092] Preferably, in the long-distance arc detection method of the above embodiments, the lower limit of the high-frequency enhancement range is... It is 20kHz.

[0093] Furthermore, in the long-distance arc detection method of the above embodiments, the upper limit of the high-frequency enhancement range... The range of 70 kHz to 100 kHz determines the maximum limit of the high-frequency oscillations and rapid transient characteristics of the electric arc that can be captured. Below 70 kHz, the high-frequency characteristic peaks of the electric arc will be truncated, and the complete frequency characteristics of the electric arc cannot be characterized. Above 100 kHz, random electromagnetic noise in the ultra-high frequency band is easily incorporated, and higher requirements are placed on the frequency used and hardware processing capabilities.

[0094] Preferably, in the long-distance arc detection method of the above embodiments, the upper limit of the high-frequency enhancement range is... It is 80kHz.

[0095] Preferably, in the long-distance arc detection method of the above embodiments, the maximum gain is set to... .

[0096] Preferably, in the long-distance arc detection method of the above embodiments, the center enhancement frequency... .

[0097] Preferably, in the long-distance arc detection method of the above embodiments, .

[0098] Optionally, in the long-distance arc detection method of any of the above embodiments, step S400 includes:

[0099] S410. Dataset preparation: Weighted random sampling is used to balance the normal samples and arc samples in the dataset for the dual-modal images, as shown in the following formula:

[0100] ,

[0101] in, Indicates sample The weight, Indicates the first The true category of each sample is 0 for normal and 1 for electric arc. Indicates batch size; This indicates that the category in the current batch of data is The number of samples;

[0102] S420. Training Preparation: Divide the dataset proportionally into training, validation, and test sets. Use the AdamW optimizer and an improved ArcFace loss function to optimize the feature space. Enhance inter-class separation by introducing angular margins, as shown in the following formula:

[0103] ,

[0104] in, This represents the value of the loss function; Indicates the first The non-true category of a sample, when the true category Normal time When the true category When it is an electric arc ; This represents the feature scaling factor, used to control the sensitivity of the loss function to the angular interval; The angular interval parameter determines the minimum angular separation between the arc and normal samples in the feature space. Indicates the first Individual samples and the true category From the perspective of weighting, Indicates the first Individual samples and non-true categories From the perspective of weighting;

[0105] S430. Model training uses a cosine annealing strategy to dynamically adjust the learning rate, as shown in the following formula:

[0106] ,

[0107] in, Indicates the current number of training steps Learning rate at that time; and These represent the minimum and maximum learning rate boundaries, respectively. Indicates the current training steps. This represents the total number of training steps;

[0108] S440. Model training is complete. Training ends when the F1 score on the validation set does not improve. The formula for calculating the F1 score is as follows:

[0109] ,

[0110] in, Indicates the current number of training steps F1 score at that time Indicates the current number of training steps The number of samples that are actually electric arcs and are correctly predicted as electric arcs by the model; Indicates the current number of training steps The number of samples that were actually normal but were incorrectly predicted as electric arcs by the model; Indicates the current number of training steps The number of samples that are actually electric arcs but were incorrectly predicted as normal by the model; set the long-distance electric arc detection model to inference state and return to step S200.

[0111] Preferably, in the long-distance arc detection method of the above embodiments, the batch size is set to... .

[0112] Optionally, in the long-distance arc detection method of any of the above embodiments, the training set: validation set: test set = 14:3:3.

[0113] Furthermore, in the long-distance arc detection method of the above embodiments, the feature scaling factor... The scaling factor ranges from 32 to 128. The lower limit of 32 provides basic gradient adjustment to ensure stable convergence of the model, which is suitable for detection tasks with high discriminative power of electric arc features. The upper limit of 128 provides a strong gradient amplification effect, which is suitable for classification tasks where the difference between electric arc and normal samples is small. However, it should be noted that an excessively large scaling factor may lead to unstable training.

[0114] Preferably, in the long-distance arc detection method of the above embodiments, the feature scaling factor With a value of 64, it provides a stable gradient signal in most arc detection scenarios, effectively promoting a compact intra-class and separated feature space distribution.

[0115] Furthermore, in the long-distance arc detection method of the above embodiments, the angle interval parameter The range is 0.3 to 0.8. The lower limit of 0.3 sets a smaller inter-class separation degree, which is suitable for scenarios where the characteristics of normal samples and arc samples are well distinguishable; the upper limit of 0.8 sets a larger inter-class separation degree, which is suitable for complex scenarios where the characteristics of normal operating conditions and arc faults are highly similar and difficult to distinguish.

[0116] Preferably, in the long-distance arc detection method of the above embodiments, the angle interval parameter With a value of 0.5, it can effectively distinguish the characteristic patterns of normal current fluctuations from those of real arc faults in arc detection tasks.

[0117] Preferably, in the long-distance arc detection method of the above embodiments, the minimum learning rate boundary is... The learning rate is 0.0001, representing the maximum boundary value. It is 0.001.

[0118] Optionally, in the long-distance arc detection method of any of the above embodiments, the learning rate starts from a maximum value of 0.001 and decreases smoothly according to the shape of a cosine function, decaying to a minimum value of 0.0001, providing a continuous and smooth learning rate change, ensuring that the model can converge quickly during training.

[0119] Optionally, in the long-distance arc detection method of any of the above embodiments, step S500 includes:

[0120] S510, receives dual-modal images;

[0121] S520, Long-distance arc detection: This function detects long-distance arcs to obtain a probability vector indicating normal operation or arc fault, i.e., the detection result. The detection result includes the arc fault confidence level and the normal operation confidence level, as shown in the following formula:

[0122] ,

[0123] in, This is the decision threshold;

[0124] S530. Execute subsequent actions. When the detection result is determined to be an arc fault, issue an alarm and execute protection actions. When the detection result is determined to be normal operation, return to step S200 and continue detection.

[0125] Furthermore, in the long-distance arc detection method of the above embodiments, the decision threshold... Set the threshold between 0.5 and 0.7, and set a lower threshold in scenarios with high security requirements.

[0126] Preferably, in the long-distance arc detection method of the above embodiments, the decision threshold... It is 0.5.

[0127] This invention designs attenuation compensation and noise enhancement strategies for long-distance transmission characteristics, improving the robustness of the model under real-world conditions. It presents time-frequency analysis and time-domain waveforms in parallel within the same image using a dual-modal imaging approach, preserving the complete information chain of time, frequency, and amplitude. Targeted enhancement is implemented in specific frequency bands of the electric arc, improving the ability to capture characteristic frequency components of the arc. A lightweight convolutional neural network structure combined with a dedicated loss function is employed to improve computational efficiency while maintaining accuracy. This invention solves the problem of arc detection under long-distance signal attenuation, providing a more accurate arc detection method.

[0128] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0129] Figure 1 This is a flowchart of an exemplary method for long-distance arc detection;

[0130] Figure 2 This is a comparison chart of the accuracy of an exemplary embodiment and the prior art;

[0131] Figure 3 This is a comparison graph of ROC curves for an exemplary embodiment and the prior art. Detailed Implementation

[0132] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0133] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of components is schematically exaggerated in some places in the drawings.

[0134] This invention designs a long-distance arc detection method, such as... Figure 1 As shown, it includes the following steps:

[0135] S100, Long-distance Arc Detection Model Construction: Based on deep learning mechanisms, a long-distance arc detection model is constructed, including an initial convolutional layer, a backbone feature layer, an attention enhancement layer, a long-distance enhancement layer, and a decision classification layer. The initial convolutional layer extracts basic features, the backbone feature layer extracts high-level visual features of the basic features, the attention enhancement layer integrates an attention mechanism to adaptively weight the high-level visual features to obtain attention-enhanced features, the long-distance enhancement layer further enhances long-distance signal transmission-related features through global pooling and fully connected layers to obtain long-distance feature-enhanced features, and the decision classification layer adopts a fully connected layer structure to transform the long-distance feature-enhanced features into decision classifications, obtaining normal / arc binary classification probabilities, and setting the state of the long-distance arc detection model to the training state; specifically including:

[0136] S110. Construct the initial convolutional layer and extract basic features using a 3×3 convolutional kernel, as shown in the following formula:

[0137] ,

[0138] in, This represents the basic features output by the initial convolutional layer. This represents the model input, i.e., the bimodal image. Indicates the convolution kernel weights; Indicates the bias term;

[0139] S120. Construct the backbone feature layer using a lightweight convolutional neural network based on the MobileNet architecture. The basic features are processed through cascaded bottleneck blocks to obtain high-level visual features, as shown in the following formula:

[0140] ,

[0141] in, Indicates high-level visual features; This indicates the number of cascaded bottleneck blocks, which is 16, allowing for sufficient feature extraction and balancing accuracy and efficiency. This represents the cascaded bottleneck block operation; the cascaded bottleneck block includes extended convolution, depthwise separable convolution, compression activation module and projection layer. Extended convolution increases the number of channels in the input data to a higher dimension, depthwise separable convolution performs feature extraction, compression activation module adaptively calibrates the channel feature response, and projection layer reduces the number of channels to the required output channels.

[0142] S130. Construct an attention enhancement layer to enhance high-level visual features and important features related to arc detection through a channel attention mechanism, i.e., attention-enhanced features, as shown in the following formula:

[0143] ,

[0144] in, Indicating attention enhancement features, This represents a multilayer perceptron. This represents the Sigmoid activation function. This indicates channel-by-channel multiplication. Indicates global average pooling. Indicates global max pooling;

[0145] S140. Construct a long-range enhancement layer. Capture the overall statistical information of the attention-enhanced features through global average pooling. Perform a non-linear transformation through two fully connected layers to extract the long-range feature enhancement features. The formula is as follows:

[0146] ,

[0147] in, Features that represent long-range feature enhancement and Let represent the weight matrices of the first transform fully connected layer and the second transform fully connected layer, respectively. and Let represent the bias vectors of the first transform fully connected layer and the second transform fully connected layer, respectively. express Activation function express Activation function;

[0148] S150. Construct a decision classification layer, including a first classification layer and a second classification layer, to transform the long-distance feature enhancement features into decision classifications, namely the confidence scores for normal state or electric arc, as shown in the following formula:

[0149] ,

[0150] in, Indicates the confidence level of a normal state or an electric arc. Indicates the confidence level of the normal state. Indicates the confidence level of the electric arc. and These represent the weight matrices of the first and second classification layers, respectively. and These represent the bias vectors of the first and second classification layers, respectively.

[0151] S160. Update the model state and put the long-distance arc detection model into training state.

[0152] S200, Signal Acquisition and Preprocessing: Raw current signals are acquired at key nodes in the power system, including combiner boxes, inverters, line interfaces, important load connection points, historically fault-prone sections, and the coordination boundaries of system protection devices. When the long-distance arc detection model is in training mode, the raw current signals are labeled as either normal samples or arc samples, with a 1:1 ratio. Normal samples are current signals acquired during normal power system operation, while arc samples are current signals acquired when an arc occurs in the power system. Arcs occur due to insulation damage, loose connections, etc., during actual operation, or triggered by an arc generator in experimental scenarios. When the long-distance arc detection model is in inference mode, the raw current signals are unlabeled. The raw current signals are standardized to obtain standardized current signals, and then preprocessed to obtain preprocessed current signals. Specifically, this includes:

[0153] S210. Raw current signal acquisition: Acquire the raw current signal at a sampling frequency of... 250kHz, signal length For 1024 sampling points, exceeding The signal is truncated, and the insufficient signal is cut off. The signal is zero-filled to ensure uniform input signal length; when the long-distance arc detection model is in the training state, the original current signal carries the label of normal sample or arc sample, and when the long-distance arc detection model is in the inference state, the original current signal does not carry the label.

[0154] S220. Standardization processing: The zero-mean, unit variance method is used to eliminate the effects of baseline drift and amplitude variation. The formula is as follows:

[0155] ,

[0156] in, This represents the value of the standardized current signal. Indicates the first The value of the original current signal at each sampling point n Values ​​range from 0 to N -1 is a positive integer; The mean of the original current signal is represented by the following calculation method: ; The standard deviation of the original current signal is represented by the following method: ;

[0157] S230, Adaptive gain compensation: Based on signal strength, adaptive gain compensation is performed on the standardized current signal to compensate for signal attenuation. The formula is as follows:

[0158] ,

[0159] in, Indicates the first The current signal value after adaptive attenuation compensation at each standardized current signal sampling point; Indicates the compensation amplification factor. Used to enhance weak electric arc signals; Indicates the weak signal strength threshold. It is used to determine whether the original signal is too weak, and thus decide whether attenuation compensation needs to be triggered;

[0160] S240. Signal Enhancement: Gaussian white noise is added to the current signal after adaptive attenuation compensation. When the long-distance arc detection model is in the training state, the intensity of the Gaussian white noise is greater than 0; when the long-distance arc detection model is in the inference state, the intensity of the Gaussian white noise is 0. The preprocessed current signal is obtained, as shown in the following formula:

[0161] ,

[0162] in, The value of the current signal that represents noise enhancement, i.e., the pre-processed current signal; Let the noise sequence have a mean of 0 and a variance of . The normal distribution Indicates the noise intensity factor. This allows for an acceptable balance between improving generalization ability and maintaining feature integrity.

[0163] S300, Dual-modal image generation: Processing the preprocessed current signal, including calculating the spectral value using short-time Fourier transform, calculating the spectral energy, enhancing the spectral energy, drawing a high-frequency enhanced spectrum, extracting key time-domain features, drawing a time-domain waveform marked with key time-domain features, and fusing the high-frequency enhanced spectrum and the time-domain waveform into a dual-modal image. If the long-distance arc detection model is in training mode, step S400 is executed; otherwise, step S500 is executed. Specifically, this includes:

[0164] S310. Calculate the frequency domain distribution. Apply a short-time Fourier transform to the preprocessed current signal to calculate the spectrum values. The formula is as follows:

[0165] ,

[0166] in, Indicates time frame and frequency point Spectral value at; This indicates the window length, which is 128 sampling points. Indicates the time frame index. This represents the total number of time frames, calculated as follows: , This indicates frame shift, preferably 32 sampling points; Indicates frequency index, ,in, ; This represents the Hanning window function, centered at... ;

[0167] S320. Calculate the spectral energy. Square the spectral values ​​to calculate the spectral energy, and then convert the linear-scale spectral energy to the logarithmic scale using a logarithmic transformation, as shown in the following formula:

[0168] ,

[0169] in, This represents the spectral energy after the logarithm of the transformation, in decibels (dB). This represents a small constant added to avoid taking the logarithm of zero. ;

[0170] S330, Spectral Energy Enhancement: Enhances spectral energy in the frequency band where arc characteristics are prominent, highlighting the high-frequency features of the arc. The formula is as follows:

[0171] ,

[0172] ,

[0173] in, Indicates at frequency point The magnitude of the increase at that location; and These represent the lower and upper limits of the high-frequency enhancement range, respectively. 20kHz 80kHz; Indicates the maximum gain. ; Indicates the center enhancement frequency, ; Indicates the enhancement parameter, ; It represents exponentiation with base e;

[0174] S340. Draw a high-frequency enhanced spectrum diagram. The high-frequency enhanced spectrum diagram is drawn in the form of a heatmap. The time axis and frequency axis are constructed as the horizontal and vertical axes, respectively, with units of milliseconds (ms) and kilohertz (kHz). The formula is as follows:

[0175] ,

[0176] ,

[0177] in, Indicates the first The actual time corresponding to each time frame; Frequency index The corresponding frequency, spectral energy After normalization, energy-color mapping is performed to the spectrogram using the Viridis color scale; key temporal features are extracted, and temporal waveforms marked with key temporal features are plotted.

[0178] S350. Extract key time-domain features: Extract key time-domain features from the preprocessed current signal, including peak position, signal RMS value, and waveform factor, as shown in the following formula:

[0179] ,

[0180] ,

[0181] ,

[0182] in, This indicates the peak position, which is the sampling position corresponding to the point where the absolute value of the signal amplitude is the largest; It represents the effective value of the signal, reflecting the average energy; This represents the waveform factor, which characterizes the peak characteristics of the signal;

[0183] S360. Draw a time-domain waveform. Based on key time-domain features, draw a time-domain waveform marked with key time-domain features, constructing a time axis as the horizontal axis, with the unit being milliseconds (ms). The formula is as follows:

[0184] ,

[0185] in Indicates the first The time value corresponding to each sampling point;

[0186] The vertical axis represents the noise-enhanced current signal value. Mark key feature parameters;

[0187] S370. Generate a dual-modal image by fusing the high-frequency enhanced spectrogram with the time-domain waveform marked with key time-domain features in a 1×2 sub-graph format to generate a dual-modal image, mining the time-frequency joint features of the signal, and improving the feature expression capability. When the long-distance arc detection model is in the training state, execute step S400; otherwise, execute step S500.

[0188] S400, Long-distance arc detection model training: Receive dual-modal images as a dataset, dynamically adjust the learning rate using a cosine annealing strategy to complete the training of the long-distance arc detection model, set the state of the long-distance arc detection model to the inference state, and return to step S200; specifically including:

[0189] S410. Dataset preparation: Weighted random sampling is used to balance the normal samples and arc samples in the dataset for the dual-modal images, as shown in the following formula:

[0190] ,

[0191] in, Indicates sample The weight, Indicates the first The true category of each sample is 0 for normal and 1 for electric arc. Indicates batch size, ; This indicates that the category in the current batch of data is The number of samples;

[0192] S420. Training Preparation: Divide the dataset into training, validation, and test sets in a ratio of 14:3:3. Use the AdamW optimizer and an improved ArcFace loss function to optimize the feature space. Introduce angular margins to enhance inter-class separation, as shown in the following formula:

[0193] ,

[0194] in, This represents the value of the loss function; Indicates the first The non-true category of a sample, when the true category Normal time When the true category When it is an electric arc ; This represents the feature scaling factor, used to control the sensitivity of the loss function to the angular interval. With a value of 64, it provides a stable gradient signal in most arc detection scenarios, effectively promoting a compact intra-class and separated inter-class feature space distribution; The angular interval parameter determines the minimum angular separation between the arc and normal samples in the feature space. ; Indicates the first Individual samples and the true category From the perspective of weighting, Indicates the first Individual samples and non-true categories From the perspective of weighting;

[0195] S430. Model training uses a cosine annealing strategy to dynamically adjust the learning rate, as shown in the following formula:

[0196] ,

[0197] in, Indicates the current training steps Learning rate at that time; and Let these represent the minimum and maximum learning rate bounds, respectively. (Minimum learning rate bounds) The value is 0.0001, and the maximum learning rate bound is... The learning rate starts at 0.001, and decreases smoothly according to the shape of a cosine function until it reaches a minimum of 0.0001, providing a continuous and smooth learning rate change and ensuring that the model can converge quickly during training. Indicates the current training step count. This represents the total number of training steps;

[0198] S440. Model training is complete. The long-distance arc detection model training is finished when the F1 score on the validation set does not improve. The formula for calculating the F1 score is as follows:

[0199] ,

[0200] in, Indicates the current training steps F1 score at that time Indicates the current training steps The number of samples that are actually electric arcs and are correctly predicted as electric arcs by the model; Indicates the current training steps The number of samples that were actually normal but were incorrectly predicted as electric arcs by the model; Indicates the current training steps The number of samples that are actually electric arcs but were incorrectly predicted as normal by the model; set the long-distance electric arc detection model to inference state and return to step S200.

[0201] S500: Long-distance arc detection, receiving dual-modal images; when an arc fault is detected, an alarm is issued; when normal operation is detected, return to step S200 to continue detection; specifically including:

[0202] S510, receives dual-modal images;

[0203] S520. Long-distance arc detection: This function detects long-distance arcs and obtains a probability vector classifying the operation as normal or arc fault, i.e., the detection result. The detection result includes the arc fault confidence score and the normal operation confidence score, as shown in the following formula:

[0204] ,

[0205] in, As the decision threshold, It is 0.5;

[0206] S530. Execute subsequent actions. When the detection result is determined to be an arc fault, issue an alarm and execute protection actions. When the detection result is determined to be normal operation, return to step S200 and continue detection.

[0207] To verify the technical effect of the present invention, an experiment was conducted. A long-distance electric arc scenario was simulated using an extension line and an electric arc generator. The electric arc generator control circuit was used to determine whether an electric arc occurred. When an electric arc occurred, an electric arc sample was obtained, and when no electric arc occurred, a normal sample was obtained.

[0208] Following step S110, an initial convolutional layer is constructed to extract basic features. Basic features are extracted using a 3×3 convolutional kernel, and the dual-modal image is processed. The RGB three-channel input is converted into a 16-channel feature representation, and the convolution kernel weights are... The bias term is Extracted basic features Following step 120, a backbone feature layer is constructed using a lightweight convolutional neural network based on the MobileNet architecture. The basic features are processed through cascaded bottleneck blocks to obtain high-level visual features. The number of cascaded bottleneck blocks is Following step S130, an attention enhancement layer is constructed to enhance high-level visual features and important features related to arc detection through a channel attention mechanism, i.e., attention-enhanced features. Following step S140, construct a long-range enhancement layer and calculate the features for long-range feature enhancement. The weight matrices of the first fully connected layer and the second transformed layer are respectively and The bias vectors of the first transform fully connected layer and the second transform fully connected layer are respectively and Following step S150, a decision classification layer is constructed, transforming the long-range feature enhancement features into decision classifications, namely the confidence levels of the normal state / electric arc. The weight matrices of the first and second classification layers are respectively... and The bias vectors for the first and second classification layers are respectively and Follow step S160 to update the model state and set the long-distance arc detection model to training state.

[0209] Following step S210, the original current signal is acquired, and the sampling frequency is... 250kHz, signal length The sampling point is set to 1024. Current signals exceeding 1024 points are truncated, and current signals with fewer than 1024 points are zero-padded to ensure a uniform input signal length. Following step S220, the zero-mean unit variance method is used to eliminate the effects of baseline drift and amplitude variation. Following step S230, adaptive gain compensation is performed on the standardized current signal based on signal strength to compensate for signal attenuation and amplification factor. Weak signal strength threshold According to step S240, when the long-distance arc detection model is in the training state, a noise intensity coefficient is added to the compensated current signal. Gaussian white noise is used to enhance anti-interference ability. When the long-distance arc detection model is in the inference state, the intensity of Gaussian white noise is 0.

[0210] Following step S310, a short-time Fourier transform is applied to the preprocessed current signal to calculate the spectral values ​​and window length. Set to 128 sampling points, frame shift Set to 32 sampling points; according to step S320, calculate the spectral energy, square the spectral values ​​to calculate the spectral energy, and then convert the linear-scale spectral energy to the logarithmic-scale using a logarithmic transformation, with a small constant. According to step S330, the spectral energy is enhanced in the frequency band where the arc characteristics are obvious, highlighting the high-frequency characteristics of the arc. The lower and upper limits of the high-frequency enhancement range are set as follows: The maximum gain is set to Center Enhancement Frequency Enhance parameters Following step S340, a high-frequency enhanced spectrum is plotted in the form of a heatmap. After the spectrum energy values ​​are normalized, energy-color mapping is performed on the spectrum using the Viridis color scale. For example, 0.0 corresponds to RGB(68,1,84) dark purple, 0.5 corresponds to RGB(42,160,121) green, and 1.0 corresponds to RGB(253,231,37). Yellow; According to step S350, extract key time-domain features from the preprocessed current signal, including peak position, signal RMS value, and waveform factors; According to step S360, based on the key time-domain features, draw a time-domain waveform diagram marked with key time-domain features, and mark the key feature parameters. The peak point corresponding to the peak position is marked with a red circle, and the RMS value is marked with a horizontal orange dashed line; the waveform factor range is mapped to the waveform curve color, smooth waveform is mapped to light blue, medium waveform is mapped to blue, sharp waveform is mapped to dark blue, and abnormally sharp waveform is mapped to a red warning color; According to step S370, fuse the high-frequency enhanced spectrum diagram and the feature-marked time-domain waveform diagram in a 1×2 sub-image format to generate a dual-modal image with an image size of 5×2.5 inches, a resolution of 70 DPI, a sub-image spacing of 0.05 inches, and an output format of RGB three-channel PNG image with a final pixel size of 350×175 pixels. When the long-distance arc detection model is in training state, execute step S410; otherwise, execute step S500.

[0211] Following step S410, the dual-modal images are subjected to weighted random sampling to balance the normal samples and arc samples in the dataset, with the batch size set to... Following step S420, in training preparation, the dataset is divided into training, validation, and test sets in a 14:3:3 ratio. The model is trained using the AdamW optimizer, and the feature space is optimized using an improved ArcFace loss function. Angle margins are introduced to enhance inter-class separation, and the feature scaling factor is set to... The angle interval parameter is set to Following step S430, during training, a cosine annealing strategy is used to adjust the learning rate, with the maximum and minimum learning rate boundaries set to... , According to step S440, when the F1 score on the validation set does not improve, the training is completed, the long-distance arc detection model state is set to the inference state, and the process returns to step S200.

[0212] Following step S510, receive the dual-modal image; following step S520, detect the long-distance arc to obtain the normal state / arc classification probability vector, i.e., the detection result, and set the decision threshold to... According to step S530, when an arc fault is detected, an alarm is issued and a protection action is performed; when normal operation is detected, return to S200 and continue detection.

[0213] The accuracy of the above embodiments and the prior art is compared in this invention, such as... Figure 2 As shown, the blue bars represent the accuracy of existing techniques using random forests, the orange bars represent the accuracy of existing techniques using support vector machines, the green bars represent the accuracy of existing techniques using logistic regression, the red bars represent the accuracy of existing techniques using XGBoost, the purple bars represent the accuracy of existing techniques using LightGBM, and the brown bars represent the accuracy of the above embodiments. It can be seen that the accuracy of the above embodiments reaches 0.978. Compared with existing techniques, the accuracy of existing techniques using random forests is 0.901, the accuracy of existing techniques using support vector machines is 0.800, the accuracy of existing techniques using logistic regression is 0.869, the accuracy of existing techniques using XGBoost is 0.877, and the accuracy of existing techniques using LightGBM is 0.895. In summary, the above embodiments have higher detection accuracy.

[0214] This invention compares the ROC curves of the above embodiments and existing technologies. The ROC curve is plotted as follows: adjusting the normal state / arc decision classification threshold, calculating the true positive rate and false positive rate of the long-distance arc detection model on the test set. The true positive rate is calculated as the proportion of arc samples correctly detected by the model out of the actual arc samples in the test samples. The false positive rate is calculated as the proportion of samples incorrectly classified as arcs by the model out of the actual arc samples. In the coordinate axis, the false positive rate is plotted on the horizontal axis, and the true positive rate on the vertical axis. Connecting the coordinates corresponding to all thresholds (false positive rate, true positive rate) forms the ROC curve. AUC represents the area under the ROC curve; the closer the value is to 1, the better the model's classification performance. Figure 3 As shown, the blue solid line represents the existing AUC curve using random forest, the orange dashed line represents the existing AUC curve using support vector machine, the green dashed line represents the existing AUC curve using logistic regression, the red dashed line represents the existing AUC curve using XGBoost, the purple dashed line represents the existing AUC curve using LightGBM, and the brown dashed line represents the AUC curve of the above embodiment. It can be seen that the AUC of the above embodiment reaches 0.995. Compared with existing technologies, the existing AUC value using random forest is 0.958, the existing AUC value using support vector machine is 0.896, the existing AUC value using logistic regression is 0.942, the existing AUC value using XGBoost is 0.958, and the existing AUC value using LightGBM is 0.958. In summary, the above embodiment better distinguishes between the "arc" and "normal" categories.

[0215] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for detecting long-distance electric arcs, characterized in that, include: S100. Construction of a long-distance arc detection model: Based on a deep learning mechanism, a long-distance arc detection model is constructed, and the state of the long-distance arc detection model is set to the training state. The long-distance arc detection model includes an initial convolutional layer, a backbone feature layer, an attention enhancement layer, a long-distance enhancement layer, and a decision classification layer. The initial convolutional layer extracts basic features, the backbone feature layer extracts high-level visual features of the basic features, the attention enhancement layer integrates an attention mechanism to adaptively weight the high-level visual features to obtain attention-enhanced features, the long-distance enhancement layer further enhances the long-distance signal transmission-related features through global pooling and fully connected layers to obtain long-distance feature-enhanced features, and the decision classification layer adopts a fully connected layer structure to transform the long-distance feature-enhanced features into decision classifications to obtain normal / arc binary classification probabilities. S200. Signal Acquisition and Preprocessing: Acquire raw current signals. When the long-distance arc detection model is in training mode, the raw current signals are labeled with normal samples or arc samples. When the long-distance arc detection model is in inference mode, the raw current signals are unlabeled. Standardize the raw current signals to obtain standardized current signals. Preprocess the standardized current signals to obtain preprocessed current signals. Specifically, this includes: S210. Acquisition of raw current signal: Acquire raw current signal; When the long-distance arc detection model is in training state, the raw current signal carries the label of normal sample or arc sample; When the long-distance arc detection model is in inference state, the raw current signal does not carry the label. S220. Standardization processing: The zero-mean unit variance method is used to eliminate the effects of baseline drift and amplitude variation, resulting in a standardized current signal. S230. Adaptive gain compensation: Based on the signal strength, adaptive gain compensation is performed on the standardized current signal to compensate for signal attenuation, as shown in the following formula: ， in, Indicates the first The current signal value after adaptive attenuation compensation at each standardized current signal sampling point; This indicates the compensation amplification factor, used to enhance weak arc signals; This represents the weak signal strength threshold, used to determine whether the original signal is too weak, and thus whether attenuation compensation needs to be triggered. S240, Signal enhancement: Gaussian white noise is added to the current signal after adaptive attenuation compensation. When the long-distance arc detection model is in the training state, the intensity of the Gaussian white noise is greater than 0. When the long-distance arc detection model is in the inference state, the intensity of the Gaussian white noise is 0, thus obtaining the preprocessed current signal. S300, Dual-modal image generation: A high-frequency enhanced spectrum diagram is plotted on the preprocessed current signal; key time-domain features are extracted; a time-domain waveform diagram marked with key time-domain features is plotted; the high-frequency enhanced spectrum diagram and the time-domain waveform diagram are fused into a dual-modal image. If the long-distance arc detection model is in a training state, step S400 is executed; otherwise, step S500 is executed. Specifically, this includes: S310. Calculate the frequency domain distribution by applying a short-time Fourier transform to the preprocessed current signal and calculating the spectrum value. S320. Calculate the spectral energy by squaring the spectral values ​​and then converting the linear-scale spectral energy to the logarithmic scale through logarithmic transformation. S330, Spectral energy enhancement: The spectral energy is enhanced in the frequency band where the arc characteristics are obvious, as shown in the following formula: ， ， in, Indicates at frequency point The magnitude of the increase at that location; and These represent the lower and upper limits of the high-frequency enhancement range, respectively. The lower limit of the high-frequency enhancement range... The upper limit of the high-frequency enhancement range is 20kHz. 80kHz; Indicates the maximum gain; Indicates the center enhancement frequency; Indicates the enhancement parameter; It represents exponentiation with base e; S340. Draw a high-frequency enhanced spectrum diagram. Draw the high-frequency enhanced spectrum diagram in the form of a heat map, and construct the time axis and frequency axis as the horizontal axis and vertical axis, respectively. S350. Extract key time-domain features: Extract key time-domain features from the preprocessed current signal, including peak position, signal RMS value, and waveform factor. S360. Draw a time-domain waveform diagram. Based on the key time-domain features, draw a time-domain waveform diagram marked with key time-domain features, and construct a time coordinate axis as the horizontal axis. S370. Generate a dual-modal image by fusing the high-frequency enhanced spectrogram with the time-domain waveform of the time-domain key feature marker. When the long-distance arc detection model is in training state, execute step S400; otherwise, execute step S500. S400, Long-distance arc detection model training: Receive the dual-modal image as a dataset, dynamically adjust the learning rate using a cosine annealing strategy to complete the training of the long-distance arc detection model, set the state of the long-distance arc detection model to the inference state, and return to step S200. S500: Long-distance arc detection. Receive the dual-modal image. When an arc fault is detected, issue an alarm; when normal operation is detected, return to step S200 and continue detection.

2. The long-distance arc detection method as described in claim 1, characterized in that, Step S100 includes: S110. Construct the initial convolutional layer and extract basic features using a 3×3 convolutional kernel; S120. Construct the backbone feature layer, using a lightweight convolutional neural network based on the MobileNet architecture. The basic features are processed by cascaded bottleneck block operations to obtain high-level visual features. S130. Construct an attention enhancement layer to enhance the high-level visual features and important features related to arc detection through a channel attention mechanism, i.e., attention enhancement features; S140. Construct a long-range enhancement layer, capture the overall statistical information of the attention enhancement features through global average pooling, and extract the long-range feature enhancement features through two fully connected layers for non-linear transformation. S150. Construct a decision classification layer, including a first classification layer and a second classification layer, and transform the long-distance feature enhancement features into decision classification, namely the confidence level of normal state or electric arc. S160. Update the model state and set the long-distance arc detection model to the training state.

3. The long-distance arc detection method as described in claim 2, characterized in that, The cascaded bottleneck block includes extended convolution, depthwise separable convolution, compression activation module and projection layer. The extended convolution increases the number of channels in the input data to a higher dimension. The depthwise separable convolution performs feature extraction. The compression activation module adaptively calibrates the channel feature response. The projection layer reduces the number of channels to the required output channels.

4. The long-distance arc detection method as described in claim 3, characterized in that, The number of cascaded bottleneck blocks ranges from 12 to 20.

5. The long-distance arc detection method as described in claim 1, characterized in that, The ratio of the normal sample to the arc sample is 1:

1.

6. The long-distance arc detection method as described in claim 1, characterized in that, Step S400 includes: S410. Data set preparation: The dual-modal images are subjected to weighted random sampling to balance the normal samples and the arc samples in the data set; S420. Training preparation: Divide the dataset into training set, validation set and test set according to the proportion. Use AdamW optimizer and adopt the improved ArcFace loss function to optimize the feature space. Enhance the inter-class separation by introducing angle margin. S430, Model training, uses cosine annealing strategy to dynamically adjust the learning rate; S440. Model training is complete. Training is completed when the F1 score on the validation set does not improve. The long-distance arc detection model is set to inference state. Return to step S200.

7. The long-distance arc detection method as described in claim 1, characterized in that, Step S500 includes: S510, Receive the dual-modal image; S520, Long-distance arc detection, detects long-distance arcs and obtains a probability vector for classifying normal operation or arc fault, i.e., the detection result; S530. Perform subsequent actions: when the detection result is determined to be an arc fault, issue an alarm and perform protection actions; when the detection result is determined to be normal operation, return to step S200 and continue detection.

Images